01 Context
Malagasy is a low-resource language: little annotated data, and speech recognition models trained and measured mostly on 15 to 30-second recordings. A voice assistant, however, receives short utterances.
Case study · AI / Voice
Malagasy speech recognition.
A model trained by Sooatek for the short utterances of a voice assistant, measured on speakers never heard in training, and published openly with its figures.
Context, constraint, solution, outcome.
02 Constraint
Measure honestly: 11,625 utterances of 2 to 8 seconds, spoken by 94 speakers absent from training, text normalized on both sides, each gap checked statistically (paired bootstrap, 95%).
03 Solution
A w2v-BERT 2.0 model fine-tuned from a public model and also trained on short utterances, a Malagasy language model, a separate model that restores casing and punctuation, and a version that runs on a CPU.
04 Outcome
- 12.7% word error rate on short utterances (with a language model; 14.3% without), against 21.5% for our previous model, a Whisper medium fine-tuned on the same data.
- 10.1% on full-length clips: no regression on complete recordings.
- 52 ms median latency per utterance on a GPU, against 259 ms for the Whisper medium.
What was delivered.
- Speech recognition model, under the CC-BY-4.0 license
- Malagasy language model
- Casing and punctuation restoration model
- CPU version, no graphics card needed
- Online demo, usable in the browser
- Public model card: protocol, measurements and limitations
Architecture, in principle.
What we take away.
- Measure in real-use conditions: on this corpus, existing models make noticeably more errors on short utterances than on long recordings.
- We publicly corrected a wrong measurement of our previous model (text not normalized): that model’s card says so and points to the new one.
Check for yourself.
Figures and method are public. Links to external pages.
- w2v-bert-malagasy model card Hugging Face: results, evaluation protocol, limitations
- Malagasy speech recognition demo Hugging Face: try it in the browser
Related offer.
Does your business have its own language, vocabulary or images? We adapt a model to your data, measured before and after.
Offer 08
6 to 10 weeksSpecialized AI model
A model adapted to your language, vocabulary or images, measured before and after: fine-tuning, speech, vision.
Included
- Data audit and preparation
- Training, evaluation on an unseen set
- Delivery for server, CPU or phone
- Measurement report
Pricing
Fixed price, on quote
GPU compute re-invoiced at actual cost
A similar project? Let’s talk.
A 30-minute scoping call to understand your context and constraints, then a written reply.